Why Engineering Teams Need an Agentic Layer, Not Just AI Chat

I like it, I use it, I see high potential. It is not perfect, but it gives good results. Sometimes, we joke with colleagues, that somebody re-invented the hot water: specification-driven development. AI, or human, in order to give good results, it needs good specification.

In some cases, AI can even be better than a usual developer. In my case, I don't know Kotlin, but I am a software developer, and my expertise was amplified in good direction, and I have a product that without AI it was not possible to do so fast, by myself. Just takes too much time and effort. What I can add here is that 'prompting' and 'stubbornness' are really important. To never give up, to disagree with the AI suggestions, to force it think for alternatives, over and over again, to the level where you feel satisfied. The will to search for the best possible solution!

It was not mentioned in the article, but a good framework that follows "AI Agentic Agile Methodology", and specification driven development is BMAD. Really helpful for brainstorming sessions. A lot of things that I have not known about and I have not thought about at all, were suggested and, consequently, added to the specification.

And in some cases, AI can be a pure evil, and amplifies in the bad direction. It happened to me with the 'marketing' phase. It just failed. I guess it is also connected with the model, but I am not sure here. If a model is good for software development, is it good for marketing strategies?

AI is not the bottleneck anymore.

Most teams already have access to powerful models, but still struggle to integrate AI into real workflows.

The problem is that software delivery is not a prompt problem—it’s a coordination problem.

This article argues that we need an agentic layer between humans, tools, and systems, instead of relying only on chat-based assistants.

It also introduces spec-driven development and a simple mental model: interaction layer + orchestration layer.

Curious how others are thinking about this beyond IDE-level AI.